VERITY
Methods

How Verity makes its numbers.

Everything on this site comes from public records. This page explains, in plain terms, what each detector looks for, how a provider ends up with a tier and a score, and how much each result can be trusted. The full technical write-up is at the bottom.

The data

Public federal and state datasets: Medicaid provider spending and enrollment for every state (238,015,729 spending rows, 2018 to 2024), Medicare enrollments and their owners, the national provider registry (9,726,865 providers), the OIG exclusion list, SAM.gov exclusions, Medicare revocations, market saturation, quality data, Census geography, state exclusion lists and state fee schedules. Every provider number is validated with its check digit, addresses are standardized, and names are parsed before anything is joined.

Detector 1: provider networks

Verity builds a graph in which hospices, home health agencies and nursing facilities are connected to the people and companies that own them, the buildings and suites they occupy, their phone numbers, their officials, and the addresses of companies already revoked or excluded. Owner records that refer to the same person are merged with a probabilistic matching model; addresses that host many unrelated providers, such as registered agents, are held out so they do not glue strangers together. Connected groups become networks.

Each network is measured on features such as how many of its companies were incorporated within one 90-day window, how many share a suite or phone, how many owners sit on three or more of them, and whether any member touches a public list. Each feature is compared with every other network using a robust z-score, which asks how far a value sits from the typical network in units that outliers cannot distort:

zi=xi−median⁡(x)1.4826⋅MAD⁡(x)z_i = \frac{x_i - \operatorname{median}(x)}{1.4826 \cdot \operatorname{MAD}(x)}

The score adds structure, list and context terms. A network is ranked only when it has three or more separate companies, at least one formed since 2021, two independent kinds of evidence, and fewer than half its members in a known chain. On the site the score is shown on a 0 to 100 scale within the ranked networks.

How often a highly ranked network already touches a public list
60%
top 10
48%
top 25
50%
top 50
55%
top 100
46%
top 250

Precision at the top of the ranking, using the structure score only and excluding chains, against public labels (revocations, exclusions, state terminations). The base rate across all networks is 36%. The p-value for the top 50 is 0.030 and for the top 250 is 5.0e-4. The top 10 alone is not statistically distinguishable from chance, at 0.108. The labels are incomplete and partly overlap the score, so these figures are a sanity check, not a measured accuracy.

Detector 2: more hours than a day holds

Medicaid pays for many services by the unit of time. Verity converts each provider's monthly billing on those codes into hours of hands-on care three ways. The lower bound needs no price at all: every claim line is worth at least one unit, so

hourslower=∑codeslinesc⋅minutesc60\text{hours}_{\text{lower}} = \frac{\sum_{\text{codes}} \text{lines}_c \cdot \text{minutes}_c}{60}

The point estimate divides dollars by a unit price, using the published state rate where one exists and a conservative estimate elsewhere. The conservative figure uses 1.5 times that price, which lowers the hours. A month is flagged only when the conservative hours per calendar day exceed 24:

hoursconservativedays in month>24\frac{\text{hours}_{\text{conservative}}}{\text{days in month}} > 24

Only codes that a clinician must deliver in person count toward an individual's hours; aide, technician and agency codes are billed under a supervising provider by design and are reported separately. Because several states let clinics bill under a supervising clinician, hours beyond a day from a single organization are tier B. Tier A needs the hours to come from three or more small billing organizations in the same month with at most 500 patients, more than 24 hours per patient per day, or a breach of a state's own daily cap. A rendering NPI with hundreds of patients a month, or whose billing organizations each carry dozens of rendering clinicians, is a medical director or supervising clinician on the claims: a supervisory umbrella, kept out of tier 2 however many organizations bill it.

Dollars in flagged months by tier, $ millions
44.60
tier A
5032.80
tier B
8488.74
tier C

Tier C is informational (over 12 hours a day, or agency volume under one clinician) and stays out of the referral queue.

What was found, by number of providers
14,076
Over 12 hours per day
1,866
More hours than a day holds
2,902
Over 16 hours per day
1,052
More hours than a day holds at a c
713
Agency volume billed under one cli
4
More than 24 hours per patient per
Calendar days or working days
denominatorproviders over 24 hours a daymonths
calendar days (used for labels)1,5763,587
working days (Mon to Fri)3,1147,074

Labels use calendar days, the more conservative choice. Dividing by working days roughly doubles the count, so both are shown.

Detector 3: paid after a list action

A provider appears on a public list with an effective date: a Medicare revocation on integrity grounds, an OIG exclusion, a SAM.gov exclusion, or a state Medicaid exclusion. Verity then looks for Medicaid payments in the months strictly after that date and before any reinstatement or the end of the re-enrollment bar. 16,357 providers are on such a list inside the data window; 391 of them were paid afterwards, $55.9M in total. Every match is by the NPI itself, and the name on the list must agree with the national registry; name conflicts are set aside.

How the providers were matched
matchtierproviders$ millions after
Exact NPI, name conflict (set aside)C550260.43
Exact NPI, name verifiedA42759.28
Exact NPI, name verifiedB6826.51
No NPI on the list; matched by name at high confidenceA67.19
No NPI on the list; matched by name at high confidenceB9845.36
$ millions paid after the action, by year of the action
2.42
2015
2.49
2016
1.32
2017
7.53
2018
2.12
2019
7.94
2020
6.12
2021
7.71
2022
6.13
2023
5.51
2024
0
2025
0
2026
How current each file is
T-MSIS provider spending, latest service month2024-12
T-MSIS enrollment segments, latest segment start2024-12-31
T-MSIS enrollment segments, latest dated segment end2026-12-31
Medicare revocations, latest effective date2026-05-27
OIG LEIE, latest exclusion date2026-08-20
SAM.gov, latest active date2026-09-04
State exclusion lists, latest action date2026-09-02
NPPES, latest deactivation date2026-08-09

"Excluded but still enrolled" is often just a stale enrollment file, so the headline counts months Medicaid actually paid, never enrollment status.

Procedures

A summary bill hides what was done; the procedure codes do not. Every provider page lists the specific codes behind its Medicaid dollars (T-MSIS, 2018 to 2024) and its Medicare Part B dollars (2024), with a description for each code. Two comparisons make a code meaningful. For Medicaid, the provider's dollars per patient-month on that code are ranked against every other provider billing the same code for six months or more:

percentilep,c=Pr⁡[ dq,c≤dp,c ]over providers q billing code c,dp,c=paidp,cpatient-monthsp,c\text{percentile}_{p,c} = \Pr\big[\, d_{q,c} \le d_{p,c} \,\big] \quad\text{over providers } q \text{ billing code } c,\qquad d_{p,c} = \frac{\text{paid}_{p,c}}{\text{patient-months}_{p,c}}

For Medicare, the provider's submitted charge divided by the allowed amount is compared with the median ratio across all providers billing that code, the way a reviewer would say "this provider bills eleven times Medicare for this code while peers bill four". A provider earns procedure points in the unified score when a large code sits in the top 5% of dollars per patient, when most of its dollars fall on code families with a documented history of abuse, or when its Medicare charge ratio is at least three times the usual ratio for the code. The points never change a tier.

Across the whole population, a code's lift is how much more often tier 1 and tier 2 providers bill it than providers in general:

liftc=nflagged,c/Nflaggednall,c/Nall\text{lift}_c = \frac{n_{\text{flagged},c} / N_{\text{flagged}}}{n_{\text{all},c} / N_{\text{all}}}
Codes that recur among tier 1 and tier 2 providers (at least 200 providers bill each)
codewhat it isflagged providersall providersliftpaid to flagged
90876Psychophysiological therapy with biofeedback 45 min27585168.4x$27.1M
S9480Intensive outpatient psychiatric services per diem141,05848.3x$13.2M
95923Testing of autonomic (sympathetic) nervous system function81,05427.7x$58K
G0483Drug test(s), definitive, utilizing (1) drug identification methods able to identify individual drugs and distinguish between structural isomers (but not necessarily stereoisomers), including, but not limited to gc/ms (any type, single or tandem) and lc/ms112,47916.2x$12.1M
93923Complete ultrasound study of arm and leg arteries92,04216.1x$198K
H0002Medicaid service code184,73613.9x$451K
H2011Crisis intervention service per 15 min154,21813.0x$8.2M
90846Family psychotherapy without patient, 50 minutes114,6188.7x$1.1M
H0031Medicaid service code229,9748.0x$1.0M
84550Uric acid level, blood84,1597.0x$60K
Q3014Telehealth originating site facility fee1911,6106.0x$1.5M
99497Advance care planning, first 30 minutes127,4065.9x$60K
90847Family psychotherapy with patient, 50 minutes2013,0445.6x$1.8M
G0439Annual wellness visit, includes a personalized prevention plan of service (pps), subsequent visit106,7855.4x$919
G2023Specimen collection for severe acute respiratory syndrome coronavirus 2 (sars-cov-2) (coronavirus disease [covid-19]), any specimen source1611,0025.3x$9.5M

Lift describes where the flagged population concentrates; it says nothing about any individual provider and is not used to flag anyone.

Factor desk for networks

Every network with three or more providers is measured on thirty factors in seven families: formation timing, ownership, addresses and contacts, list exposure, market, money, and identity and enrollment. Each factor carries its value, its percentile among all such networks in the risky direction, and a robust z-score. Six rate-of-change factors (formation velocity, share of new companies, Medicaid growth, owner association bursts, county saturation trend and ownership changes) are averaged as clipped z-scores into a momentum reading:

m=1k∑j=1kclip⁡(zj, −3, 3)m = \frac{1}{k}\sum_{j=1}^{k} \operatorname{clip}(z_j,\,-3,\,3)

with the outlook labelled rising fast at or above 1, rising from 0.3, steady, or cooling below minus 0.3. The desk is indicative: no outcome data exist yet to calibrate it, and it never changes a tier.

One score per provider

Every provider any detector reached gets a tier that says what the public record can prove: 1, on a public list and still paid afterwards, or adjudicated in a public enforcement record (sentenced, convicted, pleaded guilty or under a civil judgment); 2, more hours than a day holds across several organizations; 3, part of a ranked network that touches a public list, or charged in a public enforcement record and not yet adjudicated; 4, network structure alone or hours beyond a day under one organization; 5, worth knowing. The score starts from the tier, adds a bonus for each further strong finding that reached the provider independently, a bounded term for the dollars involved, and the procedure points described above:

score=min⁡(100,  basetier+8 min⁡(2, max⁡(nstrong−1, 0))+min⁡(9, log⁡10dollars)+procedure points)\text{score} = \min\Big(100,\; \text{base}_{\text{tier}} + 8\,\min\big(2,\ \max(n_{\text{strong}} - 1,\ 0)\big) + \min\big(9,\ \log_{10} \text{dollars}\big) + \text{procedure points}\Big)

Tier bases are 90, 75, 60, 45 and 25; the corroboration bonus is 8 per additional strong finding, capped at 16; the dollar term is capped at 9 (a billion dollars); procedure points run from 0 to 8. A strong finding is a tier-A list action, tier-A impossible volume with concurrency, or membership of a ranked network. An enforcement record counts as a strong finding only when no list action already documents the provider, because an exclusion follows the conviction it cites, and two records of one chain of evidence are not two findings.

Dollars are always the figure of the detector that set the tier, never a sum across detectors, so a weak indicator cannot lift a small case above a large one: payments in the months after a list action for tier 1, payments in the flagged months for tier 2, and Medicaid paid in 2024 for the network tiers. For an enforcement record too recent to have any Medicaid months after it (the spending file ends in December 2024), the figure is forward exposure, Medicaid paid in the last twelve observed months. Across the country: 685 providers in tier 1, 66 in tier 2, and 12 reached independently by two strong findings.

What the model does, and does not do

The referral packet is drafted from the evidence rows only; any statement that cites a record outside the list is removed and counted. The regulations named in a packet come from a fixed mapping of evidence types. The case chat can only call read-only tools over the same tables and must cite a row for every sentence. Reviewer rejections are classified to the kind of evidence that was wrong, and only that kind's weight moves; the weights stay informational until there are a few hundred reviews. Batch jobs extract state exclusion lists from PDFs and spreadsheets, review the time-code table, match list entries that lack an NPI by name, and adjudicate borderline owner matches, each with a written report.

Technical appendix, generated by the pipeline

Verity methods

How the warehouse is built, what was dropped, and the numbers behind every figure in the demo. All inputs are public federal data (see the pipeline code).

Build

In plain language. This section describes how the working database is built and what data is allowed into it. A SQL script is run by a Python driver to create a DuckDB file, and a companion CSV lists the time-based procedure codes along with the minutes each billed unit implies, a divisor for group codes, and deliberately low durations for untimed session codes so the impossible-days check understates rather than overstates. Per-diem codes carry no minutes and are kept only so day counts can be compared against patients times days in the month. Spending records are filtered to ten-digit identifiers on both provider fields and paid amounts between 0 and 50,000,000 dollars per provider-code-month, with one table further limited to codes on the time-code list. Enrollment segments with placeholder dates are dropped or treated as open-ended, and home state is assigned by the largest presence weight, which is not calendar days because segments overlap by plan and month. Exclusion screening relies on the LEIE file, the revoked list, and termination codes, since no SAM.gov extract was available.

the pipeline code is executed by the pipeline code into the pipeline code (DuckDB 1.5.5). Time-based HCPCS codes and the minutes implied by each unit are in the pipeline code; minutes per unit is the face time one billed unit implies, group divisor is the assumed number of participants for group codes, and untimed session codes use a conservative (low) duration so the impossible-days detector under-counts rather than over-counts. Per-diem codes are included with no minutes so day counts can be checked against patients times days in the month.

Filters on the T-MSIS Medicaid spending file: both NPIs must be ten digits, paid must be between 0 and 50,000,000 dollars per NPI-code-month, and (for the spend table only) the HCPCS code must be in the time-code list. Medicaid enrollment segments with placeholder start dates (1900-01-01, 1970-01-01) are dropped. A provider's home state is the state with the most enrolled days. Segments overlap (one per plan and month), so days in provider state is a presence weight for ranking states, not calendar days. Placeholder dates in the segments file (0001-01-01, 1900-01-01, 1970-01-01 starts; 9999-12-30 ends) are dropped or treated as open-ended. No SAM.gov extract is on disk; LEIE, the CMS revoked list and Medicaid termination codes are the exclusion sources.

Warehouse sanity checks (2026-09-05 12:11 ET)

In plain language. This section is a data quality check on the analysis warehouse before any findings are produced. It counts the rows in each source table and shows what was filtered out of the Medicaid spending file. After keeping only time-based procedure codes, valid provider identifiers, and payments between zero and 50 million dollars, the working spend table holds 56.7 million rows, about 1.22 million servicing providers, and 532.82 billion dollars, rising from 51.31 billion in 2018 to 91.68 billion in 2024. Large numbers of rows were removed for unusable billing or servicing identifiers, roughly 7.9 million and 10.4 million, so coverage is incomplete and any provider-level totals should be read with that in mind. Other checks show expected patterns, such as two owner types, all revoked records carrying a date, only 8,841 of 83,842 exclusion records having a provider identifier, and incorporation dates parsing for only part of the hospice, home health, and nursing facility files.

Tables in the pipeline code: chow, drops, enroll, hha, hospice, hospital, leie, medicare revoked hhs, nppes, optout, order referring, owners, ppef, provider state, revoked, saturation, saturation cbsa, saturation county, snf, spend, spend totals, timecodes

spend (time-based HCPCS only, clean NPIs, paid between 0 and 50M)

rowsservicing NPIs$ billions
56,701,8971,220,721532.82

Rows dropped from the Medicaid spending file

reasonrows
total rows238,015,729
paid_over_50M1,239
paid negative9,212
bad billing npi7,863,474
bad servicing npi10,371,226
servicing npi not 1 or 2 prefix7,648
rows with time code60,316,436

Medicaid enrollment

enrollment segmentsNPIs with a home state
49,470,2924,266,209

CMS enrollment files

hospiceHHASNFhospitalowner rowsCHOW rowsPPEF rows
5,80211,49414,4109,161613,5745,9992,978,925

Owner types (expect I and O)

typerows
I455,179
O158,395

Labels

revoked rowsrevoked with dateLEIE rowsLEIE with NPI
8,1368,13683,8428,841

Incorporation date parse rate

filerowsparsed inc date
hospice5,8024,389
hha11,4948,815
snf14,4108,607

Market saturation (typed county table)

rowsservice typesfirst periodlast periodmoratorium rows
1,030,290242020-01-01 to 2020-12-312025-01-01 to 2025-12-310

NPPES

NPIsindividualsorganizationsdeactivated
9,726,8657,415,2941,959,660370,737

Spend by year

yearrows$ billions
20186,404,06351.31
20197,205,02262.43
20207,130,42867.55
20218,778,69777.61
20229,181,57686.07
20239,655,69196.15
20248,346,42091.68

Top 15 time-based codes by dollars

hcpcsdescription$ billions
T1019Personal care services per 15 min106.49
99213Office visit established patient level 332.59
99214Office visit established patient level 429.72
T2016Habilitation residential waiver per diem27.78
H2016Comprehensive community support services per diem17.53
H2015Comprehensive community support services per 15 min16.20
S5125Attendant care services per 15 min15.72
90837Psychotherapy 60 min11.97
S5102Adult day care per diem9.34
T1020Personal care services per diem8.24
H2017Psychosocial rehabilitation services per 15 min8.19
90834Psychotherapy 45 min7.99
T1017Targeted case management each 15 min7.87
T2021Day habilitation waiver per 15 min7.57
H2019Therapeutic behavioral services per 15 min7.37

Detector 2: impossible days

Conversion. For each rendering NPI, billing NPI, HCPCS code and service month in the T-MSIS spending file, implied clinician hours are computed three ways: a rate-free lower bound (each claim line is at least one unit, so lines x minutes per unit), a point estimate (paid / rate pt x minutes per unit), and a conservative estimate (paid / rate cons with rate cons = 1.5 x rate pt). Because the data-driven rate is itself an upper bound on the unit price, both dollar-based figures understate hours; the label IMPOSSIBLE therefore means impossible under every assumption the method makes. Group codes are divided by the assumed participant count. Minutes per unit come from the pipeline code (CPT/HCPCS unit definitions; untimed session codes use the low end of the CPT time range).

Unit price estimation. T-MSIS suppresses every NPI-code-month cell with fewer than 12 claim lines or 12 patients (the smallest cell in the file has 12 of each), so no single-line payments exist and paid / lines is an average over 12 or more lines. Under full payment a line pays units x rate, so paid / lines = rate x mean units per line >= rate: the lower envelope of paid / lines across a state's cells is an upper bound on the unit price, and hours computed from it are lower bounds. The estimator is the 5th percentile of paid / lines per state, code and year after dropping cells below 20 percent of the median (stray partial payments); for codes Medicare also covers, where crossover lines pay only coinsurance, the median is used instead (session codes bill one unit per line). Where Minnesota publishes the rate (DHS-3945 January 2022 and April 2026, EIDBI billing grid January 2026, MH procedure grid) the highest published non-supervision variant of the code replaces the estimate (several programs share a code at different prices, and the highest keeps hours conservative). The conservative rate is 1.5 times the point rate.

Validation against Minnesota's published rates (signed error is positive when the estimate sits above the published unit price, which is the expected direction; units per line is the estimate divided by the published rate):

| estimator | median signed error % | median |error| % | code-years | |---|---|---|---| | p05 | 107.50 | 107.50 | 131 | | p05 trimmed | 175.90 | 175.90 | 131 | | p50 | 655.30 | 655.30 | 131 | | chosen estimator is >= 0.9 x published | | | 117 |

codeyearcellscrossoverpublishedpublished maxp05p05 trimmedp50err p05 trimmed %err p50 %units/line at p05units/line at p50
961302019441124.36124.369.3739.0470.56-68.60-43.300.310.57
961302020661124.36124.3644.2744.2799.51-64.40-20.000.360.80
9613020211821124.36124.3650.2052.49103.74-57.80-16.600.420.83
9613020221741124.36124.3650.5350.5394.39-59.40-24.100.410.76
9613020232531124.36124.3647.8348.2994.24-61.20-24.200.390.76
9613020242061124.36124.3646.3148.0289.62-61.40-27.900.390.72
96131201930185.0585.0567.9583.11157.61-2.3085.300.981.85
96131202041185.0585.0558.7158.71113.61-31.0033.600.691.34
961312021128185.0585.0572.5572.55158.54-14.7086.400.851.86
961312022140185.0585.0567.5367.53179.31-20.60110.800.792.11
961312023192185.0585.0551.7051.70141.56-39.2066.400.611.66
961312024161185.0585.0554.4954.49130.90-35.9053.900.641.54
97151202234094.8094.8091.8991.89213.12-3.10124.800.972.25
97151202332094.8094.80124.65124.65226.8431.50139.301.312.39
97151202440094.8094.80135.10135.10266.3342.50180.901.432.81
97153201979020.1820.1816.2142.1689.51109.00343.602.094.44
97153202090020.1820.1830.3031.63121.8256.80503.701.576.04
971532021405020.1820.1836.2238.71109.0591.90440.501.925.40
9715320221,390020.1820.1842.2042.5195.85110.70375.002.114.75
9715320233,571020.1820.1832.0432.3766.2060.40228.101.603.28
9715320244,117020.1820.1837.5037.5074.2085.80267.701.863.68
9715420216306.726.720.6610.2016.1751.70140.701.522.41
97154202215806.726.7210.0810.0817.7550.00164.201.502.64
97154202325506.726.7210.9510.9520.0962.90199.001.632.99
97154202411106.726.7210.4110.4125.7455.00283.101.553.83
97155201934020.1820.1841.8741.8781.01107.50301.502.074.01
97155202067020.1820.1834.4434.44103.3970.70412.401.715.12
971552021317020.1820.1828.0228.0255.2138.80173.601.392.74
971552022798020.1820.1829.7529.7555.5847.50175.501.472.75
9715520231,787020.1820.1828.8628.8648.3843.00139.801.432.40
9715520242,291020.1820.1834.3634.3658.3570.30189.201.702.89
97156202349020.1820.1831.2931.2960.0555.10197.601.552.98
971562024121020.1820.1834.2634.2656.4769.80179.801.702.80
G029920184409.1412.8114.3127.4483.37200.30812.103.009.12
G029920197109.1412.812.5415.7169.0371.80655.301.727.55
G029920208509.1412.8115.4323.5385.26157.40832.802.579.33
G029920219409.1412.818.3025.2294.86175.90937.902.7610.38
G029920225809.1412.8114.6635.6493.30290.00920.803.9010.21
G029920234609.1412.8129.2229.22103.16219.601,028.603.2011.29
G0299202458012.8112.8122.8626.8999.18109.90674.302.107.74
H20112018178040.5840.589.73113.14236.80178.80483.502.795.84
H20112019194040.5840.5812.7769.86131.7872.10224.701.723.25
H20112020155040.5840.5847.9073.98141.5382.30248.801.823.49
H20112021369040.5840.5863.1976.13129.4787.60219.001.883.19
H20112022346040.5840.5844.5363.87140.1757.40245.401.573.45
H20112023441040.5840.5828.5972.37169.4878.30317.601.784.18
H20112024385040.5840.5875.7778.26205.2292.80405.701.935.06
H20142018277014.2514.2530.6330.6358.18114.90308.302.154.08
H20142019277014.2514.2534.2935.7862.67151.10339.802.514.40
H20142020571014.2514.2524.1724.2362.1070.00335.801.704.36
H201420212,345014.2514.2527.6927.8673.3395.50414.601.965.15
H201420222,393014.2514.2529.4829.7482.31108.70477.602.095.78
H201420232,468014.2514.2530.8931.71100.32122.50604.002.237.04
H201420242,214014.2514.2530.1230.56108.35114.50660.302.147.60
H201520181,66304.5517.1771.5171.51272.421,471.705,887.2015.7259.87
H201520191,70104.5517.1771.0871.74233.351,476.805,028.5015.7751.29
H201520201,78004.5517.1771.7172.25184.341,487.903,951.4015.8840.51
H201520212,14004.5517.1751.2660.00175.681,218.703,761.0013.1938.61
H201520222,27204.5517.1751.0758.82175.601,192.703,759.3012.9338.59
H201520231,58404.5517.1752.1853.61149.741,078.203,191.1011.7832.91
H2015202490804.5517.1739.9543.75123.00861.602,603.309.6227.03
H201720181,089019.1219.1265.2167.14145.95251.20663.303.517.63
H201720191,204019.1219.1254.8855.74132.97191.50595.502.926.95
H201720201,658019.1219.1243.8844.51111.92132.80485.402.335.85
H201720213,144019.1219.1251.3953.27113.25178.60492.302.795.92
H201720223,346019.1219.1257.6160.38128.87215.80574.003.166.74
H201720233,747019.1219.1252.6464.85150.10239.20685.003.397.85
H201720243,366019.1219.1252.0363.50159.36232.10733.503.328.33
S510020185603.4511.5855.6955.6963.741,514.301,747.6016.1418.48
S510020195503.4511.5857.1557.5966.441,569.301,825.9016.6919.26
S5100202019703.4511.5828.9433.2864.89864.601,781.009.6518.81
S5100202148103.4511.5846.9946.9979.791,261.902,212.6013.6223.13
S5100202242403.4511.5862.9162.9184.301,723.602,343.4018.2424.43
S5100202342803.4511.5877.8177.8184.272,155.402,342.6022.5524.43
S5100202438604.5311.5882.5782.57107.151,722.602,265.3018.2323.65
S512020203503.767.9057.2457.2492.551,422.302,361.5015.2224.62
S512020217203.767.9039.6939.6997.05955.602,481.2010.5625.81
S512020226103.767.9043.8943.89106.321,067.302,727.6011.6728.28
S512020237603.767.9052.9054.15159.301,340.204,136.7014.4042.37
S512020246604.327.9069.5869.58230.071,510.705,225.8016.1153.26
S5130201890504.617.9017.1821.2638.49361.20734.904.618.35
S5130201988604.617.9021.2321.2941.16361.90792.904.628.93
S513020201,07104.617.9019.8819.8840.23331.20772.704.318.73
S513020211,66504.617.9020.6520.6540.04348.00768.604.488.69
S513020221,48504.617.9023.0023.1042.57401.10823.405.019.23
S513020231,46804.617.9024.6724.7846.40437.50906.505.3810.07
S513020241,34607.907.9039.6239.6273.83401.60834.605.029.35
S513520187702.577.9053.0253.02116.821,963.104,445.6020.6345.46
S5135201910302.577.9060.3360.3399.402,247.603,767.8023.4838.68
S5135202012902.577.9063.2463.24119.872,360.604,564.1024.6146.64
S5135202113702.577.9062.3562.35151.792,325.905,806.4024.2659.06
S5135202215402.577.9091.5191.51155.673,460.705,957.2035.6160.57
S5135202327702.577.9097.8997.89171.463,708.906,571.5038.0966.72
S5135202434207.907.9092.1892.18181.761,066.802,200.7011.6723.01
S515020187805.779.6452.7752.77104.99814.601,719.609.1518.20
S515020198705.779.6453.9253.92102.59834.501,677.909.3517.78
S515020207705.779.6461.4361.43125.94964.602,082.7010.6521.83
S515020217105.779.6466.7666.76152.811,056.902,548.3011.5726.48
S515020227505.779.6468.6668.66170.411,089.902,853.4011.9029.53
S515020239005.779.6458.2862.51189.42983.403,182.8010.8332.83
S515020248509.649.6497.6897.68216.84913.202,149.4010.1322.49
T100220186808.7115.890.9615.0026.4472.20203.501.723.04
T100220196208.7115.890.2619.6928.46126.10226.702.263.27
T100220206708.7115.890.375.1615.88-40.7082.400.591.82
T100220217808.7115.890.316.6131.40-24.20260.500.763.60
T100220225308.7115.890.3419.6134.94125.10301.102.254.01
T100220234708.7115.890.4714.1834.2162.80292.801.633.93
T1002202452013.2615.890.9836.9049.84178.30275.902.783.76
T100320214906.6911.933.3049.0095.00632.401,320.007.3214.20
T100320223106.6911.930.3449.0094.23632.401,308.507.3214.09
T10162018798023.1925.4621.3730.3359.3730.80156.001.312.56
T10162019988023.1925.467.1713.5054.32-41.80134.200.582.34
T101620201,221023.1925.4610.1513.1651.40-43.30121.600.572.22
T101620211,604023.1925.4611.4813.5053.59-41.80131.100.582.31
T101620221,677023.1925.4610.9212.1348.50-47.70109.100.522.09
T101620231,788023.1925.4610.8311.3648.24-51.00108.000.492.08
T101620241,569023.1925.4610.7011.7146.36-49.5099.900.502.00
T101720181,142016.6316.6373.6478.60175.24372.60953.804.7310.54
T101720191,211016.6316.6361.4877.45173.52365.70943.404.6610.43
T101720201,362016.6316.6357.4659.48149.38257.60798.303.588.98
T101720211,375016.6316.6357.1160.13152.32261.60815.903.629.16
T101720221,381016.6316.6353.5456.14160.32237.60864.103.389.64
T101720231,538016.6316.6360.5669.57226.42318.301,261.504.1813.62
T101720241,201016.6316.63101.17131.27272.83689.401,540.607.8916.41
T101920181,35304.906.9960.6360.6377.951,137.401,490.8012.3715.91
T101920191,36504.906.9958.2658.2678.741,089.001,506.9011.8916.07
T101920201,26304.906.9940.4646.4380.45847.501,541.709.4716.42
T1019202180704.906.998.429.9430.84102.80529.402.036.29
T1019202274204.906.997.1513.5041.08175.60738.402.768.38
T1019202378204.906.9912.7513.7842.11181.20759.402.818.59
T1019202467504.106.9916.4017.6845.46331.201,008.904.3111.09

Rate sources used nationally: p50 per line 22,177, p05 per line trimmed 7,266, published_MN 131, no rate bundled or zero paid 115.

Tests. Every code carries a personal service attribute: psychotherapy, E/M, prolonged services, evaluations, professional psychological testing, health-behavior, counseling, telehealth, nutrition and acupuncture must be delivered by the rendering clinician; technician, aide, personal-care, habilitation and other HCBS codes are billed under a supervising or agency NPI by design. Individual NPIs (NPPES entity type 1) are tested on personal-service hours only: IMPOSSIBLE_BY_LINE_COUNT when even the rate-free lower bound exceeds 24 hours per calendar day; IMPOSSIBLE_CONSERVATIVE_RATE when the conservative estimate exceeds 24; IMPLAUSIBLE_OVER_16H; ELEVATED_OVER_12H. Supervision-billable volume above 24 hours per day under one individual NPI is reported as UMBRELLA_VOLUME (tier B): it can be a legitimate agency structure or a ghost clinician, and only records can tell. Any NPI: IMPOSSIBLE_PER_PATIENT when conservative hours per patient exceed 24 per calendar day. Minnesota: EXCEEDS_MN_DAILY_CAP when conservative hours on capped EIDBI codes exceed cap x patients x days, i.e. more than every patient receiving the state's own daily maximum every day of the month (97153 and 0373T 8 h, 97155 6 h, 97154 4.5 h, 97156 4 h, 97151 8 h). Robust z-scores (median/MAD, 0.6745 scaling) are computed within state and NPPES taxonomy (state-only when fewer than 200 NPIs) and feed the ranking, never a label on their own.

Results. 2,661 rendering NPIs have at least one impossible month (1,904 impossible by line count alone, 1,699 with three or more impossible months), with $7,761.27M paid in those months; 613 of them are tier A (an individual whose impossible personal-service hours were billed by three or more different organizations in the same month, or an organization with more than 24 hours per patient per day); 0 Minnesota providers exceed the state's own daily cap allowance. The most common drivers are office E/M codes (99213, 99214) and psychotherapy (90837, 90791), which several states let clinics bill under a supervising physician's or psychologist's NPI; a single-organization impossibility is therefore tier B and reads as "verify the state's supervisory billing rule and pull records", not as a finding.

tierNPI-monthsNPIs$M
A3797344.60
B28,1052,5905,032.80
C146,48715,0568,488.74

Denominator sensitivity. Labels divide monthly hours by calendar days, the strictest physical bound. Dividing by Monday-to-Friday working days instead moves the counts as follows (individual NPIs, conservative personal-service hours):

denominatorNPIs over 24 h/dayNPIs over 16 h/day
calendar days (used for labels)1,5763,587
working days (Mon to Fri)3,1147,074

Growth and concentration indicator. Medicaid-only home and community-based billers (housing stabilization, EIDBI, personal care) never appear in the CMS enrollment files, so the network detector cannot see them, and services billed under the agency NPI are tested per patient. For every billing NPI and year the indicator records the dominant code, the share of dollars in it, dollars per patient-month, the year-over-year growth and whether the NPI was enumerated within three years. GROWTH_ANOMALY (tier C, informational) marks new NPIs with at least $500k in the year, 80 percent or more of dollars in one high-vector code, a first year at that size or three-fold growth, and dollars per patient-month in the top decile of their state and code. It is a queue for records review; many new providers grow quickly for legitimate reasons.

Convention caveat. In several states the T-MSIS rendering NPI is the supervising clinician or the group by convention, and telehealth and locum tenens arrangements can concentrate volume under one NPI legitimately. Every label here is a screening indicator to be checked against the state's supervisory-billing rules and the provider's records; none is a finding.

labelNPI-monthsNPIs$M in flagged months
ELEVATED_OVER_12H123,56014,0765,107.08
IMPOSSIBLE_BY_LINE_COUNT21,2081,8664,329.75
IMPLAUSIBLE_OVER_16H16,9722,9021,195.24
IMPOSSIBLE_CONSERVATIVE_RATE7,3591,052816.61
UMBRELLA_VOLUME6,6027131,611.82
IMPOSSIBLE_PER_PATIENT1245.06
stateNPIs with impossible months$M
CA6381,267.69
LA234674.35
FL177234.77
IN1602,356.96
TX132259.19
NY121561.02
NV112250.69
KY93209.13
AZ89260.26
TN72142.68
VA70101.80
AL63137.64

Minnesota EIDBI, CTSS, HSS and PCA codes (rendering NPIs, dollars, implied hours, rate source):

codeNPIs$Mimplied hours (M)rate source
H2015358730.3010.63published_MN
T1019179685.3924.51published_MN
H2014644266.184.67published_MN
971532,08962.140.77published_MN
9715590219.470.24published_MN
97151251.230.00published_MN
971542660.670.01published_MN
97156480.490.01published_MN

Tables: d2 rate, d2 rate validation, d2 implied, d2 npi month, d2 scored, d2 top; app rows in flags (detector D2). Code: the pipeline code

Detector 1: ghost networks

Graph. 31,706 enrollments (14,410 SNF, 11,494 HHA, 5,802 HOSPICE) and 466,715 owner or managing-employee rows. Nodes: providers, resolved owner persons (86,804) and organizations (18,936), building and suite-level addresses, NPPES phones, faxes, authorised officials, EINs, mailing addresses, secondary practice locations; CHOW buyer to seller edges. 270,399 nodes and 466,676 edges. 1,193 hub nodes (chains with 25+ facilities, buildings with 25+ tenants, phone numbers on 25+ records, and similar) are held out of component formation so national operators do not swallow the graph; shared buildings, mailing addresses and phones are down-weighted by 1/log2(1 + tenants).

Identity resolution. Owner persons are merged on the PECOS associate ID and on an exact key (last name, first three letters, ZIP5), then a Fellegi-Sunter model over six comparison fields (last name with Jaro-Winkler levels, first name with nickname and initial levels, middle initial, ZIP5/ZIP3, city, street number) is fitted by EM on 91,729 blocked candidate pairs (same state and last name, or same state, Soundex and first initial). Pairs are linked when the posterior match probability is at least 0.95 and at least one locational field agrees, so names alone never merge two people. EM fitted lambda = 0.0863, 24 pairs linked. Organizations merge on associate ID, on a normalized name (corporate suffixes stripped) plus state, and on token-set similarity of at least 94 within a state.

Communities. Connected components of the hub-free graph, with Leiden (RB configuration, resolution 1.0) applied to components above 120 providers: 1,960 communities with two or more providers.

Features per community. n prov by type and distinct organizations; incorporation bursts over distinct organizations formed 2019 or later (most organizations incorporated inside any 90, 180 or 365 day window; NPPES enumeration date when the incorporation date is missing); for-profit share; share of members formed since 2021; largest number of members at one building and at one suite; phone, fax, authorised-official, mailing-address and EIN sharing; owners tied to three or more members; label links (member NPIs on LEIE, SAM, Medicare revocations, state exclusion lists, Medicaid for-cause terminations, NPI deactivations; owner-name links to LEIE and SAM at high (name + ZIP5) or medium (name + state) confidence); CMS Market Saturation providers per 10k FFS beneficiaries for the dominant county and service, as a robust z on the log scale across all counties; county moratorium flag; Medicaid 2024 and all-years dollars (billing NPI) and Medicare 2023 hospice/HHA payments (PAC PUF).

Score. Each feature is converted to a robust z (median/MAD, capped at 5): structure = 1.5 z(burst 90 over distinct organizations, bursts of three or more only) + 1.5 z(address share/n) + 1.0 z(owner multi/n) + 1.5 z(new ratio) + 0.5 z(phone or official share/n) + 0.5 z(log size); the for-profit share is recorded as a feature but not scored, because nearly every hospice and home health agency in Los Angeles, Houston, Phoenix and Las Vegas is for-profit; context = 0.8 z(saturation) + 0.5 moratorium; labels = 2.0 excluded links (member NPI on LEIE or SAM, owner name on LEIE or SAM at high 1.0 or medium 0.5 confidence, same suite as an excluded or revoked entity 1.0, same building 0.5; cap 3) + 1.0 z(Medicaid for-cause terminations/n, bulk-coded states suppressed) + 1.0 revoked (cap 3) + 0.5 state exclusions (cap 3) + 0.3 deactivations (cap 3) + 1.5 enforcement adjudications (a member NPI, or an owner named as sentenced, convicted or pleaded guilty in a Department of Justice, HHS-OIG or state attorney general release, resolved by name at high confidence; cap 3) + 0.5 enforcement charges (indicted, charged or arrested, not yet adjudicated; cap 3). Enforcement releases reach the graph months before the exclusion list that will later cite them, and they reach every enrollment the named person owns. Owner counts use ownership and managing-control roles only (5 percent direct or indirect owners, managing employees, operational control, administrators); boards, officers and trustees are recorded but not scored, so hospital systems with a shared board do not look like networks. Ranked list eligibility: no chain or private-equity owner (317 communities are scored but held out: consolidation is not a ghost network), at least three distinct organizations, at least one organization formed since 2021, and at least two independent evidence families (structure, label, context). Every ranked community is a referral candidate for records review, not a finding. 141 communities are eligible.

Evaluation. The CMS enrollment files only contain providers that are still enrolled, so Medicare revocations cannot be held out as labels (only 0.0000 of communities contain a member revoked or excluded in 2023 or 2024: the revoked ones have already left the file). The structure-only score, which uses no label information, is instead evaluated against any label link (member NPI on LEIE, SAM, the revoked list, a state exclusion list or a for-cause Medicaid termination; owner name on LEIE or SAM; address shared with an excluded or revoked entity). Base rate 0.361; precision at K of the structure-only score among communities of three or more distinct organizations, chains excluded: P@10 = 0.60 (one-sided binomial p = 0.108), P@25 = 0.48 (one-sided binomial p = 0.151), P@50 = 0.50 (one-sided binomial p = 0.030), P@100 = 0.55 (one-sided binomial p = 0.000), P@250 = 0.46 (one-sided binomial p = 0.001). Read this honestly: the top-10 figure is ten items and is not statistically meaningful on its own; the label set is incomplete (it cannot contain providers that have already left the enrollment file) and it overlaps the inputs of the full risk score, so only the structure-only score is evaluated against it. 71,964 addresses of LEIE-excluded entities and revoked organizations were indexed for the address test.

statecommunities in top 200providersMedicaid 2024 $M
CA592,997168.70
TX16937254.90
FL13298472.40
OH9368179.20
IL72544.20
NV41220.00
MA413921.60
MI4867.30
AZ3641.00
LA2933.90

Top 25.

clusternhospiceHHASNFcityscoreburst90addrowner multiphoneexclrevokedMedicaid termsat zMedicaid 2024 $MMedicare 2023 $M
D1-000013030Van Nuys, CA18.7331121.500.0001.200.000.00
D1-00002142120Glendale, CA18.6259016.000.0001.200.051.89
D1-00003120120Burbank, CA17.4056113.000.0001.200.003.10
D1-00004352132Suffolk, VA15.79733720.500.0030.500.0010.06
D1-00005182160Glendale, CA15.2337017.000.0001.200.005.04
D1-000067070Glendale, CA14.7645110.500.0001.200.000.38
D1-000077525500Van Nuys, CA13.77161218316.000.0001.207.8854.85
D1-00008151041Dallas, TX13.70441311.500.000-0.100.1025.04
D1-000097424446Encino, CA13.36141411123.000.0001.202.6824.74
D1-00010335253Mission Hills, CA13.03510313.000.0001.200.6263.66
D1-000115218340North Hollywood, CA12.8367526.500.0001.200.8639.64
D1-000128930590Glendale, CA12.7917812220.500.0001.200.5044.38
D1-000137719580Glendale, CA12.761668212.500.0001.201.9950.44
D1-00014278190Van Nuys, CA12.6053616.500.0001.200.4238.73
D1-0001511934850Woodland Hills, CA12.56181412323.500.0001.200.7267.04
D1-0001610323800Van Nuys, CA12.49121211323.500.0001.200.8548.53
D1-000177920590Burbank, CA12.42109935.000.0001.200.2631.12
D1-000188515700Van Nuys, CA12.3710177215.500.0001.2010.2651.59
D1-000194232100San Antonio, TX12.37107522.000.0000.2011.6348.87
D1-000205913460Glendale, CA12.3278329.000.0001.204.3530.89
D1-000218219630Tarzana, CA12.0710106218.000.0001.203.1135.90
D1-00022362340Glendale, CA11.96511319.500.0001.200.0840.74
D1-000237333400Burbank, CA11.809714211.000.0001.204.56102.85
D1-00024294250Glendale, CA11.7744314.000.0001.200.0012.81
D1-000254316270Van Nuys, CA11.7666434.500.0001.200.2226.67

Tables: clusters (features, score, rank, summary, graph JSON), cluster members, d1 persons, d1 orgs, d1 providers. Code: the pipeline code

Detector 3: paid after a screening-trigger action

Rule. An NPI appears on a federal or state "must not be paid" list with an effective date, and Medicaid (T-MSIS provider spending, service months 2018-01 to 2024-12) shows paid claims in service months strictly after that month and, where the source gives one, before the window end (Medicare re-enrollment bar expiry, state reinstatement date). Dollars count each NPI-month once whether the NPI billed or rendered. Tier A grounds only: Medicare revocations under 42 CFR 424.535(a)(2),(3),(4),(5),(7),(8),(10),(12),(13),(14),(18),(19),(20),(22),(23); every OIG LEIE exclusion without a state waiver; California, New York and Texas Medicaid exclusion lists (rows carrying an NPI). Administrative revocations ((a)(1) noncompliance, (a)(6), (a)(9) alone, (a)(11), (a)(17), (a)(21)) are kept in the tables as tier B and excluded from the headline.

Headline. 16,357 NPIs are on a tier-A list with an effective date inside the data window; 391 of them have Medicaid claims with service months after the action, totalling $55.92M; 197 were paid in six or more months after the action ($50.65M). These are dollars paid after an action that should have triggered a state screening check under 42 CFR 455.436 (monthly LEIE, SAM and NPPES checks) and, for for-cause Medicare terminations and other states' terminations, a termination decision under 42 CFR 455.416. They are not "improper payments": a Medicare revocation is not by itself a Medicaid payment bar, appeals and reinstatements exist, and some payments may reflect claims that were later recouped. OIG's audit of providers terminated in one state and paid in others found $50.3M across 584 providers, so the order of magnitude is consistent.

sourcetierNPIs paid after$M aftermedian $ per NPImax months
NPPES_DEACTIVATEDC511231.205,352.0084
MEDICARE_REVOKEDA26030.5213,096.0067
MEDICARE_REVOKEDB6224.1125,525.0084
STATE_EXCL_CAC1423.3827,660.0046
STATE_AG_ADMINB1221.79278,578.0010
DOJ_ADMINB4612.4336,023.009
OIG_ADMINB3710.5960,981.0010
OIG_LEIEA127.7767,706.0055
STATE_EXCL_CAA817.2212,029.0052
OIG_ADJUDICATEDA16.476,465,610.0011
TMSIS_DECEASEDB204.0440,210.0081
STATE_EXCL_NYA353.457,470.0078
MEDICARE_REVOKEDC143.447,029.0033
STATE_EXCL_KYA252.7116,320.0041
STATE_EXCL_TXA62.3354,286.0033
STATE_EXCL_TXB32.2890,573.0025
STATE_EXCL_INA42.1978,698.0054
STATE_EXCL_WAC11.681,677,349.0028
STATE_EXCL_NDA11.511,506,469.0011
STATE_EXCL_WAA61.1168,330.0083
DOJ_ADJUDICATEDA40.6777,363.007
OIG_LEIEC20.30152,045.0043
DOJ_CHARGEDB40.295,536.009
STATE_AG_CHARGEDB20.25126,314.003
STATE_EXCL_KYC20.24120,054.0010
STATE_EXCL_MOA80.194,761.0069
STATE_EXCL_COA70.123,047.0038
SAM_OPMB40.1111,467.0027
STATE_EXCL_MDA50.085,888.0026
STATE_EXCL_AZA20.0631,240.002
STATE_EXCL_SCC20.0628,649.004
STATE_AG_ADJUDICATEDA10.0659,903.004
STATE_EXCL_COC10.0547,488.001
STATE_EXCL_SCA80.051,342.0032
STATE_EXCL_NYC50.046,519.003
STATE_EXCL_MIC10.0438,205.003
STATE_EXCL_TXC10.0113,635.007
STATE_EXCL_MTA10.00100.001
SAM_OPMC10.000.002
STATE_EXCL_MSA10.000.004
STATE_EXCL_AZC10.003,150.0010
STATE_EXCL_MIA10.00306.001
STATE_EXCL_MSC10.000.008
STATE_EXCL_NHA10.00444.002

Identity checks. Every NPI on every list must pass the NPI check digit (Luhn with the 80840 prefix). Rows on any list whose name shares no token with the NPPES record for that NPI are set aside as tier C, whatever the source (the California list's provider-number field can carry an employer's NPI, and a revocation can name a practice rather than the individual; 557 such rows, $260.43M, are excluded from every number above). Texas lists everyone ever excluded, so its rows use the reinstatement or eligible-to-reapply date as the window end and pre-2018 rows without either are tier B.

Match tiers. Every event in this detector carries the NPI itself, so the match is exact by identifier; the tier records how far the identity could be verified against NPPES. Rows without an NPI on the source list are handled separately by the name-matching script (the pipeline code) and never enter the headline. Counts are NPIs with Medicaid service months after the action, tier A and B lists combined, TMSIS terminations excluded.

identity matchtierNPIs paid after$M after
exact npi name conflictC550260.43
exact npi name verifiedA42759.28
exact npi name verifiedB6826.51
name match model highA67.19
name match model highB9845.36

File dates. "Excluded but still enrolled" is often an artefact of a stale enrollment file, so the headline never relies on enrollment status: it counts paid service months in T-MSIS after the action. The enrollment-segment figures below are reported separately and carry the file's own dates.

filelatest date in the file
T-MSIS provider spending, latest service month2024-12
T-MSIS enrollment segments, latest segment start2024-12-31
T-MSIS enrollment segments, latest dated segment end2026-12-31
Medicare revocations, latest effective date2026-05-27
OIG LEIE, latest exclusion date2026-08-20
SAM.gov, latest active date2026-09-04
State exclusion lists, latest action date2026-09-02
NPPES, latest deactivation date2026-08-09

Still enrolled. 2,576 NPIs revoked by Medicare (tier A) or excluded by OIG still hold an active Medicaid enrollment segment (T-MSIS status 02-06) more than 90 days after the action.

Medicaid stateNPIs
LA419
TX327
VA292
ID261
TN258
CA257
PA180
RI176
MI157
GA120

Deceased. T-MSIS status 80 (provider deceased) is applied by some states to organizations and to old records, so the test is restricted to individual NPIs whose latest status in that state is 80 and whose record starts 2015 or later. Tier A additionally requires NPPES to show the NPI deactivated. Tier A: 0 NPIs, $0.00M paid after; tier B (no NPPES corroboration): 20 NPIs, $4.04M.

Cross-state. Terminated for cause in one state (T-MSIS status 60, 65, 66, 67, 70, 72, 75, 78, 81; termination is the final status in that state; individuals or organizations enrolled in three or fewer states, so national chains with one mis-coded segment are excluded) and active in another state more than 90 days later: 6,370 NPIs (16,346 state pairs), 213 of them also on a federal or state exclusion list, 2,984 with Medicaid dollars after the termination ($1,484.68M). T-MSIS termination codes are state-coded and uneven, so this list is ranked with federally corroborated NPIs first and is presented as a screening queue, not a finding.

codereasonNPIs$M after
70TERM - LICENSE REVOKED4,705945.01
81TERM - STATE EXCLUSION/ DEBARMENT, ETC.947428.09
78TERM - ONSITE REVIEW/ PROVIDER IS NO LONGER OPERATIONAL434102.13
72TERM - MEDICARE/MEDICAID EXCLUSION2194.71
60TERM - ABUSE OF BILLING PRIVILEGES600.25
65TERM - FALSE OR MISLEADING INFORMATION103.51
66TERM - FEDERAL EXCLUSION/ DEBARMENT, ETC.40.97
75TERM - MISUSE OF BILLING NUMBER20.00

NPI deactivation. 511 NPIs deactivated in NPPES show Medicaid paid claims in later service months ($231.20M). Reported as tier B because states carry legacy identifiers.

Top 20 by dollars after the action (tier A lists).

NPIname (NPPES)typelistsfirst actionMedicaid statemonths paid afterfirstlast$ aftername agrees
1982736492WE CARE TRANSPORTATION2OIG_LEIE2010-01-20312018-012020-074,441,513.001
1962546176MATIAS CLINICAL LABORATORY INC2MEDICARE_REVOKED2018-08-31MO442018-092022-043,846,471.001
1548629520EMPIRE MEDICAL LLC2MEDICARE_REVOKED,STATE_EXCL_MD2020-07-31DE92020-082021-042,389,353.001
1679896484BLAKES BLESSING HEALTH CARE INC.2STATE_EXCL_TX2022-01-19TX332022-022024-102,118,903.001
1225242985KIUP KIM1STATE_EXCL_CA2018-12-10AZ522019-012023-082,046,769.001
1619941614HISHAM SADEK1MEDICARE_REVOKED,STATE_EXCL_IN2015-07-15IL542020-052024-112,029,469.001
1861407637HEALTHSMART PACIFIC INC2OIG_LEIE2021-04-20MD342021-052024-021,639,221.001
1447440359DOYLE'S YELLOW CHECKER CAB, INC2STATE_EXCL_ND2024-01-24MN112024-022024-121,506,469.001
1215266267ADVANCED SPINE AND PAIN CENTERS, PLLC2MEDICARE_REVOKED2021-11-19MD232021-122023-101,327,755.001
1457414286DM OPTICAL INC2STATE_EXCL_NY2016-09-22262018-012020-021,183,544.001
1194744185QUALITY HEALTHCARE MANAGEMENT INC2STATE_EXCL_KY2023-10-07RI142023-112024-121,092,507.001
1851726731INFINITY DIAGNOSTICS LABORATORY, INC2STATE_EXCL_NY,STATE_EXCL_SC,MEDICARE_REVOKED2022-10-31LA72022-112023-051,043,387.001
1891703922COMMUNITY CARE MEDICAL CLINICS INC2MEDICARE_REVOKED2020-03-02TX192021-042023-04960,939.001
1871571406MOHAMED ASWAD1OIG_LEIE,MEDICARE_REVOKED2016-01-20AZ552018-012022-07901,321.001
1407188543MERCRIS HOME HEALTH INC2MEDICARE_REVOKED2023-05-01TX172023-062024-10899,287.001
1831547868SHANONE CHATMAN-ASHLEY1MEDICARE_REVOKED,OIG_LEIE2020-10-23LA382020-112023-12883,542.001
1740478270FIRST IDEAL ENTERPRISES INC.2MEDICARE_REVOKED2018-10-01MI382018-112021-12843,035.001
1518931856LINDA WARREN-WATSON1STATE_EXCL_CA2020-10-31CA162020-112022-06828,250.001
1609064153QUEENS OPTOMETRIC CARE PLLC2MEDICARE_REVOKED2023-10-25NY112023-112024-09777,235.001
1336486448QOL COMMUNICATION SERVICES, LLC2MEDICARE_REVOKED2024-06-12MD62024-072024-12728,911.001

Tables: d3 events, d3 paid after, d3 enrolled after, d3 crossstate, d3 npi, d3 top; app rows in flags (detector D3). Code: the pipeline code

Unified provider risk score

Hierarchy. Every NPI any detector reached gets one row in provider risk with a tier, a score and the reasons. Tier 1: on a tier-A federal or state list and Medicaid service months after the action, or adjudicated in a public enforcement record (a Department of Justice release, an HHS-OIG enforcement record or a state attorney general release naming the provider as sentenced, convicted, pleaded guilty or under a civil judgment, resolved to the NPI by name at high confidence and reported as a name match). Tier 2: physically impossible personal-service volume billed by three or more small organizations in a month with at most 500 patients, more than 24 hours per patient per day, or over a state's own daily cap. Volume with hundreds of patients a month, or billed through organizations that each carry dozens of rendering clinicians, is a supervisory umbrella and stays in tier 4. Tier 3: member of an eligible provider community with a label link, or charged in a public enforcement record (indicted, charged, arrested or named in a complaint) and not yet adjudicated. Tier 4: structure only, or single-organization impossibility. Tier 5: informational. Score = tier base (90, 75, 60, 45, 25) + up to 8 procedure points (dollars per patient on a code in the top 5% of every provider billing it, most dollars on codes with a history of abuse, or Medicare charges far above the code's usual charge-to-allowed ratio) + 8 per additional strong finding that independently reached the NPI (a tier-A list action, tier-A concurrent impossible volume, or a ranked community; cap 16) + min(9, log10 dollars at risk), capped at 100. Dollars at risk is the figure of the detector that set the tier (service months after the action for tier 1, paid in flagged months for tier 2, Medicaid 2024 for the community tiers, and for an enforcement record too recent to have service months after it, Medicaid paid in the last 12 observed months, the forward exposure the v2 score uses), never a sum and never borrowed from a weaker indicator, so a single-organization volume flag cannot lift a small documented-action case above a large one. The county map sums each NPI once.

tiermeaningNPIs$M at riskreached by 2+ detectors
1adjudicated in a public enforcement record28985.560
1documented action, then payment39664.090
2impossible volume with concurrency6693.790
3network structure with a list link6,2801,156.380
3charged in a public enforcement record23775.530
4structure or single-organization volume2,5247,438.210
5informational2,2052,804.450

0 NPIs were reached by two or more detectors independently; corroboration is the strongest signal the pipeline produces and it is weighted accordingly.

Enforcement feed. Public enforcement releases (Department of Justice press releases through the DOJ API, the HHS-OIG enforcement actions listing, and the state attorney general actions it carries) are fetched daily and backfilled to January 2024, read into a schema by the model, and the parties resolved to NPIs by name at high confidence only. The spending file ends in December 2024, so an action after that has no observable months after it; the provider is still documented, and its dollars figure is forward exposure. 301 providers in this table are adjudicated in an enforcement record and 245 more are charged and not yet adjudicated (actions dated 2024-01-09 to 2026-09-03); 289 of the adjudicated providers appear on no exclusion or revocation list yet, which is the lead time the feed exists to provide. An enforcement record counts toward corroboration only when no list action already documents the provider, because an OIG exclusion follows the conviction it cites: that is one chain of evidence, not two.

Top 25 referral candidates.

rankNPInametypestatetierscoredetectors$ at riskreasons
11548629520EMPIRE MEDICAL LLC2DE1100.00['D3']2,389,353.00Listed on the Medicare revocation list since July 31, 2020; Medicaid still paid claims in 9 later months, $2,389,353 in total; Medicaid dollars per patient on code 99214 ($117 per patient-month) sit in the top 5% of every provider billing that code
21780780031DAVID SMITH1NC1100.00['D3']1,063,751.00Listed on the Medicare revocation list and the SC Medicaid exclusion list since August 30, 2021; Medicaid still paid claims in 7 later months, $1,063,751 in total; Medicaid dollars per patient on code 80307 ($557 per patient-month) sit in the top 5% of every provider billing that code
31851726731INFINITY DIAGNOSTICS LABORATORY, INC2NJ1100.00['D3']1,043,387.00Listed on the Medicare revocation list and the NY Medicaid exclusion list since October 31, 2022; Medicaid still paid claims in 7 later months, $1,043,387 in total; Medicaid dollars per patient on code U0005 ($390 per patient-month) sit in the top 5% of every provider billing that code
41992766299PATRICK CHAU1IA1100.00['D3']489,411.00Listed on the WA Medicaid exclusion list since May 26, 2006; Medicaid still paid claims in 41 later months, $489,411 in total; Medicaid dollars per patient on code T1015 ($730 per patient-month) sit in the top 5% of every provider billing that code; Medicare submitted charges on code 96127 are 3.3 times the typical charge-to-allowed ratio for that code
51851702971NEW WAVE DIAGNOSTIC RADIOLOGY PLLC2NY199.80['D3']580,310.00Listed on the Medicare revocation list since August 19, 2022; Medicaid still paid claims in 27 later months, $580,310 in total; Medicaid dollars per patient on code 93890 ($231 per patient-month) sit in the top 5% of every provider billing that code
61043256118IVAN ANTONEVICH1NY199.70['D3']460,209.00Listed on the CA Medicaid exclusion list since August 12, 2021; Medicaid still paid claims in 39 later months, $460,209 in total; Medicaid dollars per patient on code 20610 ($767 per patient-month) sit in the top 5% of every provider billing that code
71366879603PRESTIGE MEDICAL SUPPLY INC2AR199.70['D3']450,238.00Listed on the Medicare revocation list since September 14, 2023; Medicaid still paid claims in 14 later months, $450,238 in total; Medicaid dollars per patient on code A4259 ($82 per patient-month) sit in the top 5% of every provider billing that code
81154089241FAST LAB TECHNOLOGIES LLC2NY199.60['D3', 'ENF']3,896.00Listed on the SC Medicaid exclusion list since November 19, 2024; Medicaid still paid claims in 1 later months, $3,896 in total; Charged (indicted) per a Department of Justice release dated July 31, 2025, not adjudicated; Medicaid paid $1,843,694 in the last 12 observed months; Medicaid dollars per patient on code 87811 ($48 per patient-month) sit in the top 5% of every provider billing that code; Medicare submitted charges on code 87811 are 3.8 times the typical charge-to-allowed ratio for that code
91093811341MOHAMMAD HOQUE1VA199.40['D3']241,391.00Listed on the Medicare revocation list since January 21, 2021; Medicaid still paid claims in 15 later months, $241,391 in total; Medicaid dollars per patient on code 99231 ($121 per patient-month) sit in the top 5% of every provider billing that code
101083799365LAKKARAJ RAJASEKHAR1OH199.40['D3']226,119.00Listed on the Medicare revocation list since September 30, 2020; Medicaid still paid claims in 45 later months, $226,119 in total; Medicaid dollars per patient on code 99214 ($229 per patient-month) sit in the top 5% of every provider billing that code
111275288862NORTHWEST MEDICAL LAB INC2IL199.30['D3']221,263.00Listed on the Medicare revocation list since May 23, 2024; Medicaid still paid claims in 5 later months, $221,263 in total; Medicaid dollars per patient on code 87637 ($303 per patient-month) sit in the top 5% of every provider billing that code
121790709111MICHAEL ADIX1NC199.20['D3']154,645.00Listed on the Medicare revocation list since April 18, 2023; Medicaid still paid claims in 11 later months, $154,645 in total; Medicaid dollars per patient on code 74177 ($693 per patient-month) sit in the top 5% of every provider billing that code
131851353866SAMIR WAHIB1IN199.20['D3']152,666.00Listed on the NY Medicaid exclusion list and the Medicare revocation list since June 7, 2023; Medicaid still paid claims in 12 later months, $152,666 in total; Medicaid dollars per patient on code 99213 ($187 per patient-month) sit in the top 5% of every provider billing that code
141649669771SWIFT MEDICAL TRANSPORT LLC2VA199.10['D3']140,338.00Listed on the Medicare revocation list since May 13, 2019; Medicaid still paid claims in 4 later months, $140,338 in total; Medicaid dollars per patient on code A0428 ($712 per patient-month) sit in the top 5% of every provider billing that code
151992888184HASIB SARIJ1NY199.10['D3']118,713.00Listed on the Medicare revocation list since February 16, 2024; Medicaid still paid claims in 10 later months, $118,713 in total; Medicaid dollars per patient on code J3490 ($127 per patient-month) sit in the top 5% of every provider billing that code
161700808243KELLY TURNER1TX198.90['D3']88,543.00Listed on the TX Medicaid exclusion list since October 22, 2018; Medicaid still paid claims in 14 later months, $88,543 in total; Medicaid dollars per patient on code 99284 ($319 per patient-month) sit in the top 5% of every provider billing that code
171871826917BURBANK PREFERRED PROVIDER MEDICAL GROUP, INC2CA198.80['D3']67,101.00Listed on the Medicare revocation list since October 21, 2020; Medicaid still paid claims in 24 later months, $67,101 in total; Medicaid dollars per patient on code 99308 ($232 per patient-month) sit in the top 5% of every provider billing that code
181457658338MAGALY VELASQUEZ1CA198.80['D3', 'ENF']59,903.00Listed on the Medicare revocation list and a state attorney general release since July 19, 2024; Medicaid still paid claims in 4 later months, $59,903 in total; Adjudicated (sentenced) per a state attorney general release dated July 22, 2024; Medicaid dollars per patient on code D1110 ($87 per patient-month) sit in the top 5% of every provider billing that code
191528069820BRENTON WYNN1CA198.70['D3']51,770.00Listed on the CA Medicaid exclusion list and the Medicare revocation list since May 23, 2022; Medicaid still paid claims in 6 later months, $51,770 in total; Medicaid dollars per patient on code Q9966 ($37 per patient-month) sit in the top 5% of every provider billing that code
201174906432PHLEBXPRESS2CA198.70['D3']45,245.00Listed on the Medicare revocation list since January 20, 2023; Medicaid still paid claims in 14 later months, $45,245 in total; Medicaid dollars per patient on code 36415 ($44 per patient-month) sit in the top 5% of every provider billing that code
211265020051FA 1ST SOLUTIONS2IL198.60['D3']42,753.00Listed on the Medicare revocation list since March 1, 2024; Medicaid still paid claims in 7 later months, $42,753 in total; Medicaid dollars per patient on code G2023 ($212 per patient-month) sit in the top 5% of every provider billing that code
221962651836CHRISTIAN OCONNOR1NY198.50['D3']32,020.00Listed on the NY Medicaid exclusion list since October 3, 2021; Medicaid still paid claims in 19 later months, $32,020 in total; Medicaid dollars per patient on code D2335 ($532 per patient-month) sit in the top 5% of every provider billing that code
231497844757PAUL FELDMAN1NY198.40['D3']25,159.00Listed on the Medicare revocation list since January 24, 2024; Medicaid still paid claims in 3 later months, $25,159 in total; Medicaid dollars per patient on code A4657 ($398 per patient-month) sit in the top 5% of every provider billing that code
241679896484BLAKES BLESSING HEALTH CARE INC.2TX198.30['D3']2,118,903.00Listed on the TX Medicaid exclusion list since January 19, 2022; Medicaid still paid claims in 33 later months, $2,118,903 in total; 100% of Medicaid dollars are on codes with a history of abuse
251831417567ELIA TORRES1MD198.20['D2', 'ENF']14,399.00Adjudicated (sentenced) per a state attorney general release dated September 30, 2024; Medicaid paid $14,399 in the last 12 observed months; Hours beyond a day in 28 month(s) under one organization, which can be supervisory billing; records needed; Medicaid dollars per patient on code 92507 ($1584 per patient-month) sit in the top 5% of every provider billing that code; A second detector adds a weaker signal

Table: provider risk. Code: the pipeline code Every row is a referral candidate for records review, not a finding.

Entity resolution adjudication

In plain language. This section checks the rule that decides when two owner records refer to the same person. A sample of 224 owner-person pairs was reviewed: every pair the rule accepted, meaning a match probability of at least 0.95 plus agreement on location and on a real name, along with the strongest pairs the rule turned down. A language model was given the same six fields the statistical matching model uses and asked to judge each pair independently, and its verdict was compared to the rule's final decision rather than to the raw probability, because that probability is unreliable for pairs that share only a city, a state and a first initial. The model agreed with the applied decision on 94.6% of all pairs and on 100.0% of the 196 pairs it rated high confidence; among the 200 rejected pairs it called about 1% the same person, and among the 24 merged pairs it called 58% the same person. These judgments are stored for human review and do not alter the graph on their own, and pairs where the model confidently says same but the rule said different form the queue for the next round of matcher tuning.

Method. 224 owner-person pairs were adjudicated: every pair the applied merge rule accepted (posterior at or above 0.95 plus a locational agreement and real name agreement) and the highest-posterior pairs it rejected. They were judged by an automated adjudication with structured output (batch msgbatch_01We4rtEVHfz66XGSusawhHA) from the same six fields the EM model sees. The model's verdict is compared with the applied decision, not the raw posterior: the EM posterior alone is miscalibrated for pairs that agree only on city, state and a name initial, which is exactly why the rule requires a locational agreement.

Agreement with the applied merge decision: 94.6% over all adjudicated pairs, 100.0% over the 196 pairs the model rated high confidence.

pairspairsmodel says sameEM says same
rejected by the rule2000.010.00
merged by the rule240.581.00

Adjudications are stored in d1 er adjudications for human review and do not change the graph automatically; pairs where the model says same with high confidence and the EM said different are the review queue for the next matcher iteration.

Network factor desk

In plain language. This desk profiles every provider network that has at least three providers, scoring 30 factors grouped into seven families: formation timing, ownership, addresses and contacts, public list exposure, market conditions, money, and identity and enrollment. Each factor is reported three ways: its raw value, its percentile among all networks measured in the risky direction, and a robust z-score. Six of the factors track rate of change, and they are clipped, converted to robust z-scores, and averaged into a single momentum number with an outlook label of rising fast, rising, steady, or cooling. Across the 914 networks with an outlook, 388 are steady, 276 rising, 151 cooling, and 99 rising fast; among the top ten ranked networks shown, momentum z-scores run from 0.25 to 1.66. The section states that momentum is indicative only, that no outcome data yet exist to calibrate it, and that it never changes a network's tier."}

What it is. For every provider network with three or more providers, the desk measures 30 factors in seven families: formation timing (incorporation bursts, share of new companies, formation velocity, company age), ownership (owners on several providers, concentration, managing-employee share, association-date bursts, changes of ownership), addresses and contacts (shared buildings, suites, phones, officials, addresses of revoked companies, office plazas), list exposure (members and owners on public lists), market (county saturation, its three-year trend, moratorium history), money (Medicaid 2024, Medicare 2023, Medicaid growth 2022 to 2024, dollar concentration, billing share) and identity and enrollment (NPI issued close to incorporation, deactivated NPIs, multi-state footprint, chain share). Each factor carries its value, its percentile among all such networks in the risky direction, and a robust z-score.

Momentum. Six rate-of-change factors (formation velocity, new-company share, Medicaid growth, owner association bursts, saturation trend and ownership changes) are averaged as clipped robust z-scores into one momentum number with an outlook label: rising fast (z at or above 1), rising (0.3 to 1), steady, cooling. This is an indicative reading of what is moving, not a validated forecast; no outcome data exist yet to calibrate it, and it never changes a tier.

Outlook across networks: steady 388, rising 276, cooling 151, rising fast 99.

networkrankmomentum zoutlook
D1-0000110.25steady
D1-0000220.41rising
D1-0000331.28rising fast
D1-0000440.76rising
D1-0000551.15rising fast
D1-0000661.57rising fast
D1-0000771.25rising fast
D1-0000881.66rising fast
D1-0000990.71rising
D1-00010100.79rising